The coreg from nealjean

data_utils module can't be found

Hi nealjean,
thank you for the very interesting research paper and code example.
I want to rebuild the Coreg algorithm and apply it in my current university project.
But it seems, that the data_utils module for the function load_data can't be found anymore.

Is there a convenient alternative to it?

Regards,
Anton

It seems that the true labels of unlabeled training set might be used while training COREG.

I'm a phd student and I'm doing research into semi-supervised regression now. When searching about COREG, I find you have reproduced it in Github.
However, when I read function _compute_deltas below,

coreg/source/coreg.py

Lines 144 to 162 in 2ec44e2

    
               def _compute_deltas(self, L_X, L_y, h, h_temp): 
        
                   """ 
        
                   Computes the improvements in local MSE for all points in pool. 
        
                   """ 
        
                   deltas = np.zeros((self.U_X_pool.shape[0],)) 
        
                   for idx_u, x_u in enumerate(self.U_X_pool): 
        
                       # Make prediction 
        
                       x_u = x_u.reshape(1, -1) 
        
                       y_u = self.U_y_pool[idx_u].reshape(1, -1) 
        
                       y_u_hat = h.predict(x_u) 
        
                       # Compute neighbors 
        
                       omega = h.kneighbors(x_u, return_distance=False)[0] 
        
                       # Retrain regressor after adding unlabeled point 
        
                       X_temp = np.vstack((L_X, x_u)) 
        
                       y_temp = np.vstack((L_y, y_u)) 
        
                       h_temp.fit(X_temp, y_temp) 
        
                       delta = self._compute_delta(omega, L_X, L_y, h, h_temp) 
        
                       deltas[idx_u] = delta 
        
                   return deltas

I am surprised to find that you use y_u in line 152 rather than y_u_hat in line 153 to train h_temp, and using self.U_y_pool to add to training sets in line 218 of coreg.py. .

coreg/source/coreg.py

Lines 215 to 221 in 2ec44e2

    
           if max(deltas) > 0: 
        
               self.to_add['x' + str(idx_h)] = self.U_X_pool[np.argmax( 
        
                   deltas)].reshape(1, -1) 
        
               self.to_add['y' + str(idx_h)] = self.U_y_pool[np.argmax( 
        
                   deltas)].reshape(1, -1) 
        
               self.to_add['idx' + str(idx_h)] = self.U_idx_pool[np.argmax( 
        
                   deltas)]

Since both y_u and self.U_y_pool are parts of true labels of unlabeled training set self.y_unlabeled or self.U_y,

coreg/source/coreg.py

Line 152 in 2ec44e2

y_u = self.U_y_pool[idx_u].reshape(1, -1)

coreg/source/coreg.py

Lines 241 to 242 in 2ec44e2

    
           self.U_X_pool, self.U_y_pool, self.U_idx_pool = shuffle( 
        
               self.U_X, self.U_y, range(self.U_y.size))

and the variable y_u_hat which stores the estimated labels of unlabeled training set is mentioned only once in the file, I am doubtful whether you use the true labels of unlabeled training set in training process. In my opinion, unlabeled y shouldn't be used in training in semi-supervised learning task.

nealjean / coreg Goto Github PK

coreg's People

Stargazers

Watchers

Forkers

coreg's Issues

data_utils module can't be found

It seems that the true labels of unlabeled training set might be used while training COREG.

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	def _compute_deltas(self, L_X, L_y, h, h_temp):
	"""
	Computes the improvements in local MSE for all points in pool.
	"""
	deltas = np.zeros((self.U_X_pool.shape[0],))
	for idx_u, x_u in enumerate(self.U_X_pool):
	# Make prediction
	x_u = x_u.reshape(1, -1)
	y_u = self.U_y_pool[idx_u].reshape(1, -1)
	y_u_hat = h.predict(x_u)
	# Compute neighbors
	omega = h.kneighbors(x_u, return_distance=False)[0]
	# Retrain regressor after adding unlabeled point
	X_temp = np.vstack((L_X, x_u))
	y_temp = np.vstack((L_y, y_u))
	h_temp.fit(X_temp, y_temp)
	delta = self._compute_delta(omega, L_X, L_y, h, h_temp)
	deltas[idx_u] = delta
	return deltas

	if max(deltas) > 0:
	self.to_add['x' + str(idx_h)] = self.U_X_pool[np.argmax(
	deltas)].reshape(1, -1)
	self.to_add['y' + str(idx_h)] = self.U_y_pool[np.argmax(
	deltas)].reshape(1, -1)
	self.to_add['idx' + str(idx_h)] = self.U_idx_pool[np.argmax(
	deltas)]

	self.U_X_pool, self.U_y_pool, self.U_idx_pool = shuffle(
	self.U_X, self.U_y, range(self.U_y.size))